What Stocks Would a Neural Network Buy?

SMITH BRAIN TRUST – Ever wonder how the Terminator would do picking stocks? Researchers at the University of Maryland's Robert H. Smith School of Business might have the answer. Working with 900,000 financial news articles and a neural network, the researchers found they could predict stock returns for up to 13 weeks.

Studies have long shown that share prices quickly respond to stock splits, earnings releases and other company news. Historically, prices responded to numerical information, such as revenue data or earnings per share.

"And yet most information is textual. It's words," says Smith School finance professor Steve Heston, who authored the study with Smith PhD graduate Nitish R. Sinha. "It's newspaper articles and financial press wires."

And that led Heston and Sinha to ask the question: Is it possible to process those data and measure the effect on stock prices?

Their research, "News Versus Sentiment: Predicting Stock Returns from News Stories," adds to the well-established body of research about efficient markets and is forthcoming in the Financial Analysts Journal. The research employed a dataset of 900,754 news articles from the Thomson Reuters news system, a proprietary Thomson-Reuters neural network and the work of a computational linguist.

The neural network effectively "reads" the articles and identifies the subject, object and verb of each sentence, taking into account intensifying words, such as "very" or "many" and negations, such as "not," and then "rates" how positive or negative the article is article based on a probability scale.

"We used it to show that we could use it to buy companies with positive news and sell companies with negative news and beat the market," says Heston.

The researchers found that the duration of stock return predictability depended, in part, on the temporal aggregation of news stories. When news was measured for one day, stock predictability lasted for just a few days. But when news was measured for one week predictability could last for up to 13 weeks, or until the next earnings release, whichever comes first.

They also found that, generally, bad news might have a longer-term impact than good news.

One possible explanation for that, says Heston, is that when there is good news about a company, any market participant, from amateur to professional, can buy up shares. However, when there's bad news, only short-sellers and the people who own the stock can sell it. And that's a far smaller proportion of the overall pool of market participants.

"There are a lot more people who can take action on good news than bad news," Heston says. "And so we find that good news is reflected in the stock price in a week or two, but bad news is what takes up to 13 weeks to be reflected."

Adverse news often doesn't become fully priced in by the market until the next scheduled earnings announcement, Heston says. He says that also suggests that investors, digesting bad news, might be waiting for a slip in revenue or other confirmation of the effect of the bad news before selling off shares.

"We always say the market is efficient -- that the stock market quickly and efficiently processes all available information so that the stock price today is the best predictor of the stock price tomorrow," Heston says.

There likely are other plausible explanations for the longer stock return predictability after a bad-news event, Heston acknowledges. That's one of the curious outtakes of this research.

"And the puzzle from the efficient-markets perspective is: Why doesn't the stock go down on the day of the announcement?" he says.

Or, alternatively, why doesn't every trader buy a neural network that can digest thousands of financial news articles a day?

"If you were running an asset management firm or a hedge fund, this tells you that you need to find news and interpret it in more than a superficial manner," Heston says.

And likely, he says, that's just step one. In a few years, he predicts, major market players may be using neural networks to evaluate Facebook posts, message board posts, tweets and, potentially, facial expressions on cable business news networks.

"This is a burgeoning field for research," he says.

For Heston, the findings also underscore the Smith faculty's drive to be on the forefront of research. He recently presented his findings at a gathering of elite investors and industry experts at the Institute for Quantitative Research in Finance, known as the Q Group.

"You could join billionaire investors at the Q Group to learn this," he says. "Or you could come here to the Smith School and get it first."